A Research Agenda on Pediatric Chest X-Ray: Is Deep Learning Still in Childhood?

Abstract

Several reasons explain the significant role that chest X-rays play on supporting clinical analysis and early disease detection in pediatric patients, such as low cost, high resolution, low radiation levels, and high availability. In the last decade, Deep Learning (DL) has been given special attention from the computer-aided diagnosis research community, outperforming the state of the art of many techniques, including those applied to pediatric chest X-rays (PCXR). Due to this increasing interest, much high-quality secondary research has also arisen, overviewing machine learning and DL algorithms on medical imaging and PCXR, in particular. However, these secondary studies follow different guidelines, hampering their reproduction or improvement by third-parties regarding the identified trends and gaps. This paper proposes a "deep radiography" of primary research on DL techniques applied in PCXR images. We elaborated on a Systematic Literature Mapping (SLM) protocol, including automatic search on six sources for studies published from January 1, 2010, to May 20, 2020, and selection criteria utilized on a hundred research papers. As a result, this paper categorizes twenty-six relevant studies and provides a research agenda highlighting limitations, gaps, and trends for further investigations on DL usage in PCXR images. Besides the fact that there is no systematic mapping study on this research topic, to the best of authors' knowledge, this work organizes the process of finding and selecting relevant studies and data gathering and synthesis in a reproducible way.